Q and A with Anton de Weger, AEP Digital and Technology Lead

Tell us about your role at the AEP?

As the AEP Technology and Digital Lead, I use technology to create automated processes around the key functions for the AEP, which includes providing a system that manages the participants data from the different testing they undergo and generate a report back to the referring neurologist, as well as make sure that the data collected can support future research.  

This technology covers a lot of areas including designing the approach, selecting the appropriate software and systems, working with technology vendors, building systems to collect and securely store the data, removing identifiable information, integration between multiple systems, automation of manual processes and using machine learning to process and analyse data. 

All these processes will help the AEP scale from tens and hundreds of participants to thousands or more.
‍

How did the portal come about?

We needed a user-friendly secure method to deliver back the results of the study to the treating neurologist. The portal is where all the measurement data is filtered down to the most important components and included in a standardised AEP Report.
‍

When it first launched what features did it have?

An initial review of the software market didn’t find anything that was fit-for-purpose at an acceptable price, so it was decided that the best approach would be to build our own clinician portal. 

This has been delivered through an iterative approach, so the first version was pretty limited, only covering user authentication, report viewing and download functions, new report notification via email, the ability to search for historic patients and make new referrals.
‍

How has it evolved? 

There has been a lot of new functionality added since the first release. This includes the ability to view patient details, mobile device support, AEP Report generation automation (which saves hours of AEP team members’ time per report), the ability for neurologists to access other participants’ information in an emergency and grouping of patients by clinic.

A comprehensive MRI viewer has also been added to the portal, allowing neurologists to quickly access the high-resolution structural MRIs that are collected as part of the study without having to navigate multiple hospital systems and different viewers.

Additionally, new workflows have also been added for expressions of interest to join the AEP from people with epilepsy and control group volunteers, along with a pathway to refer your patient for neurologists. 
‍

What features does the AEP Report have? 

The AEP Report covers the main components a neurologist would consider when treating their patient; the radiologist report, genetics report, cognitive testing results, language processing analysis and hippocampal volume analysis.
‍

What is unique about these features and what value do they provide referring clinicians? Anyone else?

The clinical and medical research experts of the AEP have selected the key factors that assist in determining the diagnosis and treatment of epilepsy. The AEP Report brings these together in one place. However, this is just the starting point, the real value of the AEP study is the standardised data that is collected using the latest technology and approaches. This standardised data allows the AEP researchers to identify similarities across participants and compare them to the study’s control group volunteers. This is expected to lead to much better understanding of how epilepsy impacts different people and will ultimately lead to much more insightful information being added to the AEP Report.
‍

How is machine learning used for the AEP report?

The volume of data and complexity of the brain means that it is extremely difficult to see patterns across the different cohort data sets, therefore advanced machine learning techniques are required to find these patterns and create predictions that can then be used by the neurologist to determine next steps for their patient. Current and future uses of machine learning include:

• Image filtering and quality assurance: Training up models that improve the quality of the images that highlight the features linked to epilepsy and detecting issues with images like movement while being scanned.
• Assisting the radiologist: Using identified items in other participants’ scans to highlight areas of interest in new scans.  
• Genetics: Correlating genetic similarities across participants and linking them; best medications, side effects, image features, cognitive testing results and outcomes.  We collect over 600,000 data points per participant from genetics, so this is a big job.
• Automation of manual processes: Using the historic data to speed up the easy to detect anomalies while still pushing through the less clear cases to experts for analysis. This includes things like working out which side of the brain is being used to process language, which can be impacted by epilepsy.
• Process improvement:  We can use machine learning to predict which participants are at higher risk and prioritise their testing.  Another use is in the understanding of the large amount of medical history documentation, which in many cases are low resolution faxes, and identify key points that need to be recorded, like date of first seizure, medication prescriptions, etc.
• Clinician support:  Providing clinicians with analytical data about how their patient compares to study participants, which may help with diagnosis, selection of best medication, etc.  We can also use Generative AI to assist with a neurologist’s workflow; like automatic generation of a draft letter back to the GP, summarising the patient’s condition.
• Research: Looking for correlations across the multimodal data (e.g. images, genetics, cognitive testing results, medical history and other study data). This could provide exciting new avenues for research and better understanding of epilepsy.

‍

What factors do you need to mitigate against when applying machine learning (ML) and artificial intelligence (AI) to healthcare?

Understanding how machine learning works is key to being able to build good healthcare applications. Firstly, it is critical that machine learning is used as a tool that helps the neurologist understand their patient, and the decision making remains the neurologist’s responsibility.  

In many cases, this means that the reasons behind why a prediction is made is as important as the prediction from the machine learning algorithm. This focus on explainability helps us better understand the prediction and validate that it is reasonable and not an anomaly. 

There is also the issue of unwanted bias in machine learning trained algorithms, where the training data might introduce patterns that can obscure the information that is being sought. An example could be an imbalance of genders, age, education, or other demographics preventing machine learning algorithms from working correctly for patients that aren’t strongly represented in the training data. This is a concern when making predictions on extremely complex interactions like in brain function.

There are many other concerns like understanding the confidence level of predictions, how they impact and how this prediction can be used, regulation requirements and preserving participant anonymity in machine learning models.
‍

How do we protect participant and volunteer data?

All the participant and volunteer data is stored in Australia using secure cloud computing data centres and processes. Only de-identified information is shared with researchers and collaborators. All data is encrypted both in storage and transmission through the internet. Multifactor identification is used in systems that interact with participant data.
‍

What are the future plans for the portal/platform?

This is the most exciting part about the AEP. We are building a world-leading, state of the art, epilepsy dataset, but it is the knowledge that we gain from this dataset that will be the most impactful. It will lead to a better understanding of epilepsy and how the brain works. It is the inclusion of technology within the research project that allows for the scaling of the impact, analysis of research data and improved outcomes.

In the short term we are working on a Medical History Explorer that allows us to search through historic medical documentation for a participant, specifically focused on answering important questions for the study. Things like; Is there a family history of epilepsy? When was the first seizure? What medications have been prescribed? What were the symptoms of the seizures, and so forth.

Another interesting use of technology that we are working on is a participant chatbot that is trained on the best epilepsy research and knowledge as well as the AEP study process, to be a helpful assistant answering participants’, volunteers’, or anyone’s questions about epilepsy in a way that aligns with the AEP clinical and medical experts’ views and can be trusted.

In the longer term we are looking to solve the big questions for epilepsy such as “Will I have another seizure?”, “Am I a good surgery candidate?”, “What medication would work best for me?” and “Can I get my driver’s licence back?”

‍